Researcher Jiahao Shao has alleged that Anthropic flagged his work as training a competing model and then blocked, downgraded or interrupted Claude sessions involving AI training. His account, circulated on October 31, 2026, says an open-source model supplied the training rollouts, while Claude helped submit computing jobs and analyze results.
The allegation remains unconfirmed. The reviewed sources contain no response from Anthropic addressing Shao’s specific case, and they do not provide the underlying conversation logs or enforcement notice needed to determine what triggered the restrictions. His account nevertheless drew substantial attention on Hacker News, where readers debated research access and dependence on a single AI provider.
Anthropic’s published documentation supplies an important distinction: its safeguards cover both distillation attempts and some frontier-model development tasks. That means the practical question extends beyond whether Claude’s outputs became training data. It also concerns which kinds of assistance Anthropic permits around the training process.
What Shao Says Happened
Shao said Anthropic tagged his work as “training a competitor,” despite his assertion that he had not distilled Claude. According to the post reproduced by conv.news, the affected work included:
- Reinforcement fine-tuning, or RFT: training that improves a model using reward signals.
- Supervised fine-tuning, or SFT: training on examples of desired inputs and outputs.
- Larger-model training: a broader category whose precise scope was not established in the account.
Shao also alleged that Claude Opus 5.5 was downgraded to Opus 5, or that sessions were terminated. These are his descriptions of the behavior, not independently verified product findings.
His central distinction was about Claude’s role in the workflow. He said the rollouts came from a strong open-source model and that “Claude just helped submit jobs and run analysis.”
A rollout is a generated response or sequence of model actions that can be used during training or evaluation. Its origin matters: using Claude’s responses as training targets is a different activity from asking Claude to prepare a job submission or interpret an experiment’s logs.
Shao’s allegation concerns Claude assistance around training work. It should not be read as evidence that Anthropic was hosting his fine-tuning jobs or directly shutting down external computing infrastructure.
Hacker News Reaction Focused on Dependence and Fallbacks
The opened Hacker News discussion showed 299 points and 169 comments. Conv.news separately displayed an engagement indicator of roughly 1,300 for the circulated post. Those are different counters captured at different stages of circulation, not figures to combine into one audience total.
Several commenters interpreted the allegation as a reason to reduce dependence on Anthropic. Hacker News user zmmmmm recommended “actively cultivating a set of alternatives and fallbacks to use,” arguing for more portable professional workflows. Another commenter, Bender, said they had already moved one project elsewhere because of perceived restrictions.
The discussion also contained disagreement about how much preparation was worthwhile. User arjie argued for using the most capable model available and moving to alternatives when guardrails intervene, questioning the value of maintaining techniques that newer models may make obsolete.
These comments document the reaction within that thread. They are not independent confirmation of Shao’s experience or a representative survey of Claude customers.
My view is that the portability concern stands apart from the accusation. A research pipeline can become fragile whenever an external assistant helps operate it but can change models or refuse a request mid-session. That is an architectural dependency worth planning for, even before the cause of any individual restriction is settled.
Anthropic Distinguishes Distillation From Illicit Extraction
Anthropic’s published anti-distillation account explicitly acknowledges that the technique itself has legitimate uses:
“Distillation is a widely used and legitimate training method.”
In ordinary model distillation, a student model learns from a stronger teacher’s outputs. A lab might use that approach to produce a smaller, cheaper version of its own model.
Anthropic says the campaigns it opposes involve unauthorized extraction of Claude’s capabilities. In its account, DeepSeek, Moonshot and MiniMax generated more than 16 million exchanges through approximately 24,000 fraudulent accounts, violating its terms and regional access restrictions. Those figures and attributions are Anthropic’s claims, not independent measurements presented here.
The company says it identified coordinated activity using request metadata, IP correlations, infrastructure indicators and, in some cases, corroboration from industry partners. It also describes uses beyond copying final answers, including generating reasoning material and using Claude to grade outputs as a reward model for reinforcement learning.
That last detail matters when assessing a training workflow. An open-source model can generate the rollouts while another model supplies judgments that influence training. Rollout provenance alone therefore cannot describe every possible form of capability transfer.
Shao said Claude’s role was job submission and analysis. The available record does not establish whether any Claude output entered a training dataset, reward calculation or optimization process.
The Consequential Detail: Training Assistance Has Its Own Trigger
Anthropic’s model-switching support page lists distillation attacks and certain frontier-model development tasks as separate safeguard categories.
The latter includes distributed training infrastructure, machine-learning accelerator design and kernel development for certain non-standard chips. Anthropic describes these blocking safeguards as “intentionally broad.”
That separation is the most useful policy detail for interpreting Shao’s complaint. A workflow could avoid using Claude as a training-data source yet still encounter a restriction because of the development assistance being requested.
The fit is partial, however:
| Shao’s reported experience | What Anthropic documents | What remains unresolved |
|---|---|---|
| Restrictions around RFT, SFT and larger-model training | Safeguards for a limited category of frontier-model development work | The page does not say all fine-tuning or training analysis is blocked |
| Automatic model downgrades | Classifier-triggered switching to another Claude model | The documented page covers Fable 5 and Fable 5.1, not his claimed Opus 5.5 route |
| Interrupted or blocked sessions | Blocking when switching is disabled, plus separate safeguards on fallback models | The specific cause of his reported interruptions is unknown |
The version mismatch matters. Shao described Opus 5.5 switching to Opus 5. The support page explains Fable 5 and Fable 5.1 falling back to Opus models. It documents a relevant mechanism and policy category, but cannot verify the model transition he reported.
The supported inference is narrower than the strongest claims in the discussion: Anthropic has published a reason why some training-related assistance may be restricted independently of distillation. That makes the distinction between those categories important in any explanation of Shao’s case.
I would want an enforcement notice to identify which category applied. A warning about extracting model capabilities and a restriction on frontier-training infrastructure describe different problems. Researchers need that distinction to understand whether they can change the workflow, request review or move the task elsewhere.
How Context Can Keep Triggering a Fallback
Frequently Asked Questions
3 questions
1Has Anthropic Confirmed Shao’s Allegation?
No confirmation or response to his specific account appears in the reviewed sources. The allegation remains unconfirmed.
